Cvpr2022 unified upgrade of oral | cosface and arcface, and adaface solves face recognition of low-quality images

all the time , Face recognition of low-quality images is challenging , Because face attributes are blurred and degraded .margin-based loss functions The progress of has improved the recognizability of human face in embedded space . Besides , Previous studies have studied the impact of adaptive loss , Classify errors （Head） The sample is more important .

In this work , Another factor is introduced into the loss function , namely image quality . The author thinks that , It is emphasized that the strategy of misclassifying samples should be adjusted according to their image quality . say concretely , The relative importance of simple or difficult samples should be given based on the image quality of the samples . Therefore, the author proposes a new loss function to emphasize the importance of different difficult samples through image quality .

The method of this paper is through feature norms To approximate image quality , This is achieved in the form of adaptive edge function .

A lot of experiments show that ,AdaFace stay 4 Data sets (IJB-B、IJB-C、IJB-S and IJBTinyFace) It improves the existing (SoTA) Face recognition performance .

Thesis link ：

https://arxiv.org/abs/2204.00964

Main contributions ：

1. A loss function is proposed , AdaFace, It gives different weights to different difficult samples according to the image quality of samples . adopt Combined with image quality , Avoid emphasizing images that are difficult to recognize , Focus on difficult but identifiable samples ;
2. Experiments show that , The learning gradient of corner edge scale is related to the difficulty of training samples . This observation prompted the author to emphasize difficult samples by adaptively changing the edge function , If the image quality is low , Ignore very difficult samples （ Unrecognized image ）.
3. Proved feature norms It can be used as an agent of image quality . It bypasses the need for an additional module to estimate image quality . therefore , Adaptive marginal functions do not require additional complexity .
4. Through to 9 Data sets of different quality (LFW、CFP-FP、CPLFW、AgeDB、CALFW、IJB-B、IJB-C、IJB-S and TinyFace) Extensive assessment of , The validity of this method is verified . Experiments show that , AdaFace The recognition performance on low-quality data sets can be greatly improved , While maintaining performance on high-quality data sets .